El apartado a del ejercicio 3 es una continuación del ejercicio 2, donde deberás analizar la técnica de transfer learning sobre el dataset de mascarillas. Por tanto, necesitarás hacer uso de lo visto en la práctica 5.3.
Empleando el dataset de mascarillas del ejercicio 2, debes probar lo siguiente:
La entrega de este ejercicio se realiza a través de la tarea creada para tal efecto en Enseñanza Virtual. Tienes que entregar un notebook, y el HTML generado a partir de él, cuyas celdas estén ya evaluadas.
La estructura del notebook debe contener los siguientes apartados:
HONESTIDAD ACADÉMICA Y COPIAS: un trabajo práctico es un examen, por lo que debe realizarse de manera individual. La discusión y el intercambio de información de carácter general con los compañeros se permite (e incluso se recomienda), pero NO AL NIVEL DE CÓDIGO. Igualmente el remitir código de terceros, OBTENIDO A TRAVÉS DE LA RED o cualquier otro medio, se considerará plagio.
Cualquier plagio o compartición de código que se detecte significará automáticamente la calificación de CERO EN LA ASIGNATURA para TODOS los alumnos involucrados. Por tanto a estos alumnos NO se les conservará, para futuras convocatorias, ninguna nota que hubiesen obtenido hasta el momento. SIN PERJUICIO DE OTRAS MEDIDAS DE CARÁCTER DISCIPLINARIO QUE SE PUDIERAN TOMAR.
Para la carga de de datos subiremos el conjunto de dataset como un zip a google drive y lo descomprimiremos desde ahí
from google.colab import drive
drive.mount('/content/drive')
Mounted at /content/drive
# Copiamos el dataset y lo decomprimimos
!cp /content/drive/MyDrive/mask_ds.zip .
!unzip -q mask_ds.zip
# Desmontamos
drive.flush_and_unmount()
Como ya hemos visto en las prácticas, cuando se trabaja con un dataset pequeño, es muy aconsejable usar redes ya entrenadas, para reutilizar ese aprendizaje obtenido en ese dataset mas grande en el nuestro, aún habiendo sido entrenado con imagenes muy distintas.
Para los problemas la elección de un modelo es importante ya que una transferencia de aprendizaje adecuada acelera el proceso de aprendizaje y ayuda a nuestro modelo a que generalice
Para este primer modelo he optado por usar VGG16, este es un modelo utilizado para clasificación de imágenes (como es nuestro caso).
He realizado algunas pruebas con imágenes sueltas y la tasa de acierto con usuarios con mascarilla es bastante alta(aunque detecta diferentes tipos de máscara, entiende bien que la persona de la foto lleva algún tipo de máscara en la cara)
El clasificador que he decidido usar en los 3 modelos ha sido la primera opción vista en la práctica (aplicar la base convolucional sobre el dataset y esos datos aplicarlos a un clasificador independiente)
he optado por este ya que es el que menos recursos consume de los dos ya que mi ordenador le lleva bastante tiempo procesar la información
from tensorflow.keras.applications.vgg16 import VGG16, decode_predictions
modeloVGG16 = VGG16()
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg16/vgg16_weights_tf_dim_ordering_tf_kernels.h5 553467096/553467096 [==============================] - 5s 0us/step
from tensorflow.keras.utils import plot_model
conv_base_VGG16 = VGG16(weights='imagenet',
include_top=False,
input_shape=(150, 150, 3))
plot_model(conv_base_VGG16, to_file='conv_base_plot.png', show_shapes=True, show_layer_names=True)
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg16/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5 58889256/58889256 [==============================] - 0s 0us/step
Aqui lo que vamos a hacer es transfer learning de los modelos preentrenamos que hemos elegido, es decir, vamos a usar la red ya preentrenada sobre muchos datos para extraer las características relevantes. He usado el ejemplo dado en la práctica 5.3
import os
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
base_dir = 'mask_ds'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
test_dir = os.path.join(base_dir, 'test')
datagen = ImageDataGenerator(rescale=1./255)
batch_size = 20
def extract_features(directory, sample_count):
features = np.zeros(shape=(sample_count, 4, 4, 512))
labels = np.zeros(shape=(sample_count))
generator = datagen.flow_from_directory(
directory,
target_size=(150, 150),
batch_size=batch_size,
class_mode='binary')
i = 0
for inputs_batch, labels_batch in generator:
features_batch = conv_base_VGG16.predict(inputs_batch)
features[i * batch_size : (i + 1) * batch_size] = features_batch
labels[i * batch_size : (i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
# sin el break estaria cargando indefinidamente
break
return features, labels
train_features, train_labels = extract_features(train_dir, 2000)
validation_features, validation_labels = extract_features(validation_dir, 1000)
test_features, test_labels = extract_features(test_dir, 1000)
train_features = np.reshape(train_features, (2000, 4 * 4 * 512))
validation_features = np.reshape(validation_features, (1000, 4 * 4 * 512))
test_features=np.reshape(test_features, (1000, 4 * 4 * 512))
Aquí lo que he hecho ha sido aplanar las características para poder usarlas en una capa densa en mi modelo
Para este segundo modelo he optado por usar VGG19, al igual que con el modelo anterior he probado con varias imágenes del conjunto y detecta bien las mascarillas (le da un alto porcentaje a la probailidad de que tengan una máscara).
El modelo VGG19 es un modelo entrenado para clasificación y he estado investigando por internet y fue muy usado (y se sigue usando) durante la pandemia para detectar si la gente usaba mascarilla,por lo que lo he usado este modelo preentrenado como segunda opción
from tensorflow.keras.applications.vgg19 import VGG19
modeloVGG19 = VGG19()
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg19/vgg19_weights_tf_dim_ordering_tf_kernels.h5 574710816/574710816 [==============================] - 7s 0us/step
conv_base_VGG19 = VGG19(weights='imagenet',
include_top=False,
input_shape=(150, 150, 3))
plot_model(conv_base_VGG19, to_file='conv_base_plot2.png', show_shapes=True, show_layer_names=True)
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg19/vgg19_weights_tf_dim_ordering_tf_kernels_notop.h5 80134624/80134624 [==============================] - 0s 0us/step
def extract_features_VGG19(directory, sample_count):
features = np.zeros(shape=(sample_count, 4, 4, 512))
labels = np.zeros(shape=(sample_count))
generator = datagen.flow_from_directory(
directory,
target_size=(150, 150),
batch_size=batch_size,
class_mode='binary')
i = 0
for inputs_batch, labels_batch in generator:
features_batch = conv_base_VGG19.predict(inputs_batch)
features[i * batch_size : (i + 1) * batch_size] = features_batch
labels[i * batch_size : (i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
# sin el break estaria infinitamente cargando valores
break
return features, labels
train_features_VGG19, train_labels_VGG19 = extract_features_VGG19(train_dir, 2000)
validation_features_VGG19, validation_labels_VGG19 = extract_features_VGG19(validation_dir, 1000)
test_features_VGG19, test_labels_VGG19 = extract_features_VGG19(test_dir, 1000)
Found 4433 images belonging to 2 classes. 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 9s 9s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step Found 2267 images belonging to 2 classes. 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 10s 10s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 8s 8s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step Found 1960 images belonging to 2 classes. 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 7s 7s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step
train_features_VGG19 = np.reshape(train_features_VGG19, (2000, 4 * 4 * 512))
validation_features_VGG19 = np.reshape(validation_features_VGG19, (1000, 4 * 4 * 512))
test_features_VGG19 = np.reshape(test_features_VGG19, (1000, 4 * 4 * 512))
Para el tercer y último modelo he utilizado EfficientNetB0, este modelo lo he usado porque en teoría nos piden usar un modelo que requiera de poca potencia(menos parámetros que los otros).
En este modelo podemos ver que el número de capas es mucho mayor a los dos anteriores, en lo que se basa este modelo es en encontrar un conjunto óptimo que minimice los recursos utilizados mientras maximiza el rendimiento
from tensorflow.keras.applications import EfficientNetB0
modelEfficientNetB0 = EfficientNetB0()
Downloading data from https://storage.googleapis.com/keras-applications/efficientnetb0.h5 21834768/21834768 [==============================] - 0s 0us/step
conv_base_Eff = EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=(150, 150, 3))
plot_model(conv_base_Eff, to_file='conv_base_plot3.png', show_shapes=True, show_layer_names=True)
Downloading data from https://storage.googleapis.com/keras-applications/efficientnetb0_notop.h5 16705208/16705208 [==============================] - 0s 0us/step
def extract_features_Eff(directory, sample_count):
features = np.zeros(shape=(sample_count, 5,5 , 1280))
labels = np.zeros(shape=(sample_count))
generator = datagen.flow_from_directory(
directory,
target_size=(150, 150),
batch_size=batch_size,
class_mode='binary')
i = 0
for inputs_batch, labels_batch in generator:
features_batch = conv_base_Eff.predict(inputs_batch)
features[i * batch_size : (i + 1) * batch_size] = features_batch
labels[i * batch_size : (i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
break
return features, labels
train_features_Eff, train_labels_Eff = extract_features_Eff(train_dir, 2000)
validation_features_Eff, validation_labels_Eff = extract_features_Eff(validation_dir, 1000)
test_features_Eff, test_labels_Eff = extract_features_Eff(test_dir, 1000)
Found 4433 images belonging to 2 classes. 1/1 [==============================] - 2s 2s/step 1/1 [==============================] - 0s 484ms/step 1/1 [==============================] - 1s 504ms/step 1/1 [==============================] - 0s 473ms/step 1/1 [==============================] - 0s 430ms/step 1/1 [==============================] - 0s 434ms/step 1/1 [==============================] - 0s 481ms/step 1/1 [==============================] - 1s 754ms/step 1/1 [==============================] - 1s 769ms/step 1/1 [==============================] - 1s 787ms/step 1/1 [==============================] - 1s 560ms/step 1/1 [==============================] - 0s 488ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 0s 475ms/step 1/1 [==============================] - 0s 444ms/step 1/1 [==============================] - 0s 447ms/step 1/1 [==============================] - 0s 451ms/step 1/1 [==============================] - 1s 506ms/step 1/1 [==============================] - 0s 472ms/step 1/1 [==============================] - 0s 464ms/step 1/1 [==============================] - 0s 469ms/step 1/1 [==============================] - 0s 464ms/step 1/1 [==============================] - 0s 484ms/step 1/1 [==============================] - 0s 446ms/step 1/1 [==============================] - 0s 439ms/step 1/1 [==============================] - 0s 434ms/step 1/1 [==============================] - 1s 718ms/step 1/1 [==============================] - 1s 742ms/step 1/1 [==============================] - 1s 717ms/step 1/1 [==============================] - 1s 726ms/step 1/1 [==============================] - 0s 491ms/step 1/1 [==============================] - 0s 428ms/step 1/1 [==============================] - 0s 498ms/step 1/1 [==============================] - 0s 470ms/step 1/1 [==============================] - 0s 479ms/step 1/1 [==============================] - 0s 491ms/step 1/1 [==============================] - 0s 497ms/step 1/1 [==============================] - 0s 490ms/step 1/1 [==============================] - 0s 486ms/step 1/1 [==============================] - 0s 474ms/step 1/1 [==============================] - 1s 504ms/step 1/1 [==============================] - 0s 432ms/step 1/1 [==============================] - 0s 497ms/step 1/1 [==============================] - 0s 468ms/step 1/1 [==============================] - 1s 501ms/step 1/1 [==============================] - 1s 528ms/step 1/1 [==============================] - 1s 730ms/step 1/1 [==============================] - 1s 869ms/step 1/1 [==============================] - 1s 633ms/step 1/1 [==============================] - 0s 478ms/step 1/1 [==============================] - 0s 475ms/step 1/1 [==============================] - 0s 427ms/step 1/1 [==============================] - 0s 472ms/step 1/1 [==============================] - 0s 487ms/step 1/1 [==============================] - 1s 530ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 0s 483ms/step 1/1 [==============================] - 0s 490ms/step 1/1 [==============================] - 0s 463ms/step 1/1 [==============================] - 0s 444ms/step 1/1 [==============================] - 0s 448ms/step 1/1 [==============================] - 0s 491ms/step 1/1 [==============================] - 0s 480ms/step 1/1 [==============================] - 0s 498ms/step 1/1 [==============================] - 1s 767ms/step 1/1 [==============================] - 1s 795ms/step 1/1 [==============================] - 1s 796ms/step 1/1 [==============================] - 1s 787ms/step 1/1 [==============================] - 0s 462ms/step 1/1 [==============================] - 0s 486ms/step 1/1 [==============================] - 0s 479ms/step 1/1 [==============================] - 0s 482ms/step 1/1 [==============================] - 0s 457ms/step 1/1 [==============================] - 0s 458ms/step 1/1 [==============================] - 0s 485ms/step 1/1 [==============================] - 0s 490ms/step 1/1 [==============================] - 0s 483ms/step 1/1 [==============================] - 0s 489ms/step 1/1 [==============================] - 0s 476ms/step 1/1 [==============================] - 0s 494ms/step 1/1 [==============================] - 1s 501ms/step 1/1 [==============================] - 0s 497ms/step 1/1 [==============================] - 1s 591ms/step 1/1 [==============================] - 1s 803ms/step 1/1 [==============================] - 1s 787ms/step 1/1 [==============================] - 1s 626ms/step 1/1 [==============================] - 0s 481ms/step 1/1 [==============================] - 0s 467ms/step 1/1 [==============================] - 0s 468ms/step 1/1 [==============================] - 0s 488ms/step 1/1 [==============================] - 0s 482ms/step 1/1 [==============================] - 0s 471ms/step 1/1 [==============================] - 0s 488ms/step 1/1 [==============================] - 0s 481ms/step 1/1 [==============================] - 0s 485ms/step 1/1 [==============================] - 1s 507ms/step 1/1 [==============================] - 0s 491ms/step 1/1 [==============================] - 1s 542ms/step 1/1 [==============================] - 1s 506ms/step 1/1 [==============================] - 0s 500ms/step Found 2267 images belonging to 2 classes. 1/1 [==============================] - 1s 584ms/step 1/1 [==============================] - 1s 796ms/step 1/1 [==============================] - 1s 770ms/step 1/1 [==============================] - 1s 818ms/step 1/1 [==============================] - 1s 681ms/step 1/1 [==============================] - 1s 569ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 0s 491ms/step 1/1 [==============================] - 0s 493ms/step 1/1 [==============================] - 1s 502ms/step 1/1 [==============================] - 0s 480ms/step 1/1 [==============================] - 0s 463ms/step 1/1 [==============================] - 1s 522ms/step 1/1 [==============================] - 0s 477ms/step 1/1 [==============================] - 1s 530ms/step 1/1 [==============================] - 0s 478ms/step 1/1 [==============================] - 1s 513ms/step 1/1 [==============================] - 0s 498ms/step 1/1 [==============================] - 1s 510ms/step 1/1 [==============================] - 1s 733ms/step 1/1 [==============================] - 1s 835ms/step 1/1 [==============================] - 1s 820ms/step 1/1 [==============================] - 1s 801ms/step 1/1 [==============================] - 1s 502ms/step 1/1 [==============================] - 1s 520ms/step 1/1 [==============================] - 1s 512ms/step 1/1 [==============================] - 0s 489ms/step 1/1 [==============================] - 0s 479ms/step 1/1 [==============================] - 0s 479ms/step 1/1 [==============================] - 1s 514ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 0s 492ms/step 1/1 [==============================] - 0s 476ms/step 1/1 [==============================] - 0s 484ms/step 1/1 [==============================] - 1s 552ms/step 1/1 [==============================] - 0s 469ms/step 1/1 [==============================] - 1s 532ms/step 1/1 [==============================] - 1s 647ms/step 1/1 [==============================] - 1s 781ms/step 1/1 [==============================] - 1s 782ms/step 1/1 [==============================] - 1s 570ms/step 1/1 [==============================] - 1s 510ms/step 1/1 [==============================] - 0s 472ms/step 1/1 [==============================] - 1s 512ms/step 1/1 [==============================] - 0s 490ms/step 1/1 [==============================] - 1s 509ms/step 1/1 [==============================] - 0s 484ms/step 1/1 [==============================] - 1s 519ms/step 1/1 [==============================] - 0s 477ms/step 1/1 [==============================] - 1s 501ms/step Found 1960 images belonging to 2 classes. 1/1 [==============================] - 0s 489ms/step 1/1 [==============================] - 0s 483ms/step 1/1 [==============================] - 0s 474ms/step 1/1 [==============================] - 0s 497ms/step 1/1 [==============================] - 0s 500ms/step 1/1 [==============================] - 1s 825ms/step 1/1 [==============================] - 1s 772ms/step 1/1 [==============================] - 1s 754ms/step 1/1 [==============================] - 1s 742ms/step 1/1 [==============================] - 1s 526ms/step 1/1 [==============================] - 0s 476ms/step 1/1 [==============================] - 1s 505ms/step 1/1 [==============================] - 0s 469ms/step 1/1 [==============================] - 1s 508ms/step 1/1 [==============================] - 0s 478ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 0s 494ms/step 1/1 [==============================] - 1s 517ms/step 1/1 [==============================] - 0s 480ms/step 1/1 [==============================] - 1s 522ms/step 1/1 [==============================] - 1s 520ms/step 1/1 [==============================] - 1s 510ms/step 1/1 [==============================] - 0s 475ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 1s 524ms/step 1/1 [==============================] - 1s 707ms/step 1/1 [==============================] - 1s 763ms/step 1/1 [==============================] - 1s 893ms/step 1/1 [==============================] - 1s 626ms/step 1/1 [==============================] - 0s 488ms/step 1/1 [==============================] - 0s 496ms/step 1/1 [==============================] - 1s 531ms/step 1/1 [==============================] - 1s 505ms/step 1/1 [==============================] - 1s 505ms/step 1/1 [==============================] - 0s 495ms/step 1/1 [==============================] - 0s 484ms/step 1/1 [==============================] - 0s 483ms/step 1/1 [==============================] - 0s 469ms/step 1/1 [==============================] - 0s 490ms/step 1/1 [==============================] - 0s 477ms/step 1/1 [==============================] - 0s 468ms/step 1/1 [==============================] - 0s 477ms/step 1/1 [==============================] - 0s 486ms/step 1/1 [==============================] - 1s 590ms/step 1/1 [==============================] - 1s 765ms/step 1/1 [==============================] - 1s 877ms/step 1/1 [==============================] - 1s 829ms/step 1/1 [==============================] - 1s 761ms/step 1/1 [==============================] - 1s 623ms/step 1/1 [==============================] - 1s 520ms/step
train_features_Eff = np.reshape(train_features_Eff, (2000, 5 *5 * 1280))
validation_features_Eff = np.reshape(validation_features_Eff, (1000, 5 * 5 * 1280))
test_features_Eff = np.reshape(test_features_Eff, (1000, 5 * 5 * 1280))
En este primer modelo podemos ver que tenemos un acierto muy bueno tanto en el conjunto de entrenamiento como en el de validación (prácticamente al 100% en entrenamiento y cerca del 98 en validación).Se puede observar que conforme avanzas las epocas, se sobreajusta un poco.
He añadido dos capas con 512 y 256 neuronas con activación relu con el objetivo de que al añadir más neuronas, al disponer de un conjunto de datos no excesivamente grande quizás esta configuración ha provocado el sobreajuste que se puede ver en la gráfica
from keras import models
from keras import layers
from keras import optimizers
model = models.Sequential()
model.add(layers.Dense(512, activation='relu', input_dim = 4 * 4 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(256, activation='relu'))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))
model.compile(optimizer=optimizers.RMSprop(lr=1e-3),
loss='binary_crossentropy',
metrics=['acc'])
history = model.fit(train_features, train_labels,
epochs=30,
batch_size=20,
validation_data=(validation_features, validation_labels))
WARNING:absl:`lr` is deprecated in Keras optimizer, please use `learning_rate` or use the legacy optimizer, e.g.,tf.keras.optimizers.legacy.RMSprop.
Epoch 1/30 100/100 [==============================] - 5s 45ms/step - loss: 0.6469 - acc: 0.8150 - val_loss: 0.3163 - val_acc: 0.8690 Epoch 2/30 100/100 [==============================] - 5s 47ms/step - loss: 0.1863 - acc: 0.9335 - val_loss: 0.3722 - val_acc: 0.8670 Epoch 3/30 100/100 [==============================] - 5s 51ms/step - loss: 0.1383 - acc: 0.9555 - val_loss: 0.1552 - val_acc: 0.9330 Epoch 4/30 100/100 [==============================] - 4s 41ms/step - loss: 0.1211 - acc: 0.9595 - val_loss: 0.2237 - val_acc: 0.9230 Epoch 5/30 100/100 [==============================] - 5s 45ms/step - loss: 0.1054 - acc: 0.9685 - val_loss: 0.0774 - val_acc: 0.9670 Epoch 6/30 100/100 [==============================] - 6s 55ms/step - loss: 0.0851 - acc: 0.9735 - val_loss: 0.0995 - val_acc: 0.9660 Epoch 7/30 100/100 [==============================] - 4s 42ms/step - loss: 0.0700 - acc: 0.9780 - val_loss: 0.1901 - val_acc: 0.9470 Epoch 8/30 100/100 [==============================] - 5s 51ms/step - loss: 0.0698 - acc: 0.9755 - val_loss: 0.0990 - val_acc: 0.9580 Epoch 9/30 100/100 [==============================] - 7s 71ms/step - loss: 0.0592 - acc: 0.9820 - val_loss: 0.1908 - val_acc: 0.9490 Epoch 10/30 100/100 [==============================] - 5s 50ms/step - loss: 0.0447 - acc: 0.9875 - val_loss: 0.1404 - val_acc: 0.9680 Epoch 11/30 100/100 [==============================] - 7s 66ms/step - loss: 0.0424 - acc: 0.9875 - val_loss: 0.1118 - val_acc: 0.9760 Epoch 12/30 100/100 [==============================] - 5s 49ms/step - loss: 0.0340 - acc: 0.9905 - val_loss: 0.1218 - val_acc: 0.9660 Epoch 13/30 100/100 [==============================] - 5s 53ms/step - loss: 0.0434 - acc: 0.9895 - val_loss: 0.5347 - val_acc: 0.9140 Epoch 14/30 100/100 [==============================] - 7s 75ms/step - loss: 0.0373 - acc: 0.9905 - val_loss: 0.2641 - val_acc: 0.9580 Epoch 15/30 100/100 [==============================] - 5s 46ms/step - loss: 0.0495 - acc: 0.9860 - val_loss: 0.1037 - val_acc: 0.9700 Epoch 16/30 100/100 [==============================] - 5s 52ms/step - loss: 0.0218 - acc: 0.9940 - val_loss: 0.2967 - val_acc: 0.9490 Epoch 17/30 100/100 [==============================] - 5s 51ms/step - loss: 0.0228 - acc: 0.9955 - val_loss: 0.2278 - val_acc: 0.9550 Epoch 18/30 100/100 [==============================] - 4s 44ms/step - loss: 0.0339 - acc: 0.9945 - val_loss: 0.2092 - val_acc: 0.9630 Epoch 19/30 100/100 [==============================] - 5s 51ms/step - loss: 0.0107 - acc: 0.9940 - val_loss: 0.1710 - val_acc: 0.9660 Epoch 20/30 100/100 [==============================] - 5s 46ms/step - loss: 0.0168 - acc: 0.9960 - val_loss: 0.1787 - val_acc: 0.9660 Epoch 21/30 100/100 [==============================] - 4s 43ms/step - loss: 0.0032 - acc: 0.9990 - val_loss: 0.2399 - val_acc: 0.9560 Epoch 22/30 100/100 [==============================] - 5s 51ms/step - loss: 0.0319 - acc: 0.9925 - val_loss: 0.2246 - val_acc: 0.9610 Epoch 23/30 100/100 [==============================] - 5s 49ms/step - loss: 0.0347 - acc: 0.9950 - val_loss: 0.5995 - val_acc: 0.9370 Epoch 24/30 100/100 [==============================] - 4s 43ms/step - loss: 0.0172 - acc: 0.9955 - val_loss: 0.2458 - val_acc: 0.9590 Epoch 25/30 100/100 [==============================] - 6s 57ms/step - loss: 0.0101 - acc: 0.9965 - val_loss: 0.3053 - val_acc: 0.9490 Epoch 26/30 100/100 [==============================] - 6s 58ms/step - loss: 0.0137 - acc: 0.9965 - val_loss: 0.5817 - val_acc: 0.9460 Epoch 27/30 100/100 [==============================] - 5s 48ms/step - loss: 0.0488 - acc: 0.9900 - val_loss: 0.2703 - val_acc: 0.9580 Epoch 28/30 100/100 [==============================] - 6s 62ms/step - loss: 0.0268 - acc: 0.9940 - val_loss: 0.1830 - val_acc: 0.9650 Epoch 29/30 100/100 [==============================] - 5s 54ms/step - loss: 0.0058 - acc: 0.9975 - val_loss: 0.2008 - val_acc: 0.9690 Epoch 30/30 100/100 [==============================] - 6s 56ms/step - loss: 0.0073 - acc: 0.9975 - val_loss: 0.2812 - val_acc: 0.9670
import matplotlib
%matplotlib inline
import matplotlib.pyplot as plt
ent_acc = history.history['acc']
val_acc = history.history['val_acc']
ent_loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(ent_acc))
plt.plot(epochs, ent_acc, 'yo', label='Entrenamiento')
plt.plot(epochs, val_acc, 'y', label='Validación')
plt.title('Accuracy en Entrenamiento y Validación')
plt.legend()
plt.figure()
plt.plot(epochs, ent_loss, 'yo', label='Entrenamiento')
plt.plot(epochs, val_loss, 'y', label='Validación')
plt.title('Pérdida en Entrenamiento y Validación')
plt.legend()
plt.show()
Este segundo modelo es usando VGG19, en este podemos ver que desde las primeras epocas de adapta mejor y que aunque disminuye minimamente el rendimiento (alrededor de un 1%) tiene un poco menos de sobreajuste y es mas suave la curva.
La pérdida tambien en ambos casos es menor y esta la curva mas suavizada(en el anterior es más irregular)
Podemos ver aqui comparado con el primero que con un número menor de neuronas y capas, el modelo en validación desde el principio es mas regular.
from keras import models
from keras import layers
from keras import optimizers
model2 = models.Sequential()
model2.add(layers.Dense(256, activation='relu'))
model2.add(layers.Dropout(0.5))
model2.add(layers.Dense(1, activation='sigmoid'))
model2.compile(optimizer=optimizers.RMSprop(learning_rate=2e-5),
loss='binary_crossentropy',
metrics=['acc'])
history2 = model2.fit(train_features_VGG19, train_labels_VGG19,
epochs=30,
batch_size=50,
validation_data=(validation_features_VGG19, validation_labels_VGG19))
Epoch 1/30 40/40 [==============================] - 2s 33ms/step - loss: 0.5162 - acc: 0.7385 - val_loss: 0.4637 - val_acc: 0.7550 Epoch 2/30 40/40 [==============================] - 1s 36ms/step - loss: 0.3435 - acc: 0.8635 - val_loss: 0.3828 - val_acc: 0.8000 Epoch 3/30 40/40 [==============================] - 2s 42ms/step - loss: 0.2679 - acc: 0.8960 - val_loss: 0.3104 - val_acc: 0.8580 Epoch 4/30 40/40 [==============================] - 2s 40ms/step - loss: 0.2239 - acc: 0.9210 - val_loss: 0.2449 - val_acc: 0.8960 Epoch 5/30 40/40 [==============================] - 1s 31ms/step - loss: 0.1967 - acc: 0.9270 - val_loss: 0.2243 - val_acc: 0.9080 Epoch 6/30 40/40 [==============================] - 1s 28ms/step - loss: 0.1683 - acc: 0.9345 - val_loss: 0.2262 - val_acc: 0.9020 Epoch 7/30 40/40 [==============================] - 1s 27ms/step - loss: 0.1516 - acc: 0.9460 - val_loss: 0.2258 - val_acc: 0.8980 Epoch 8/30 40/40 [==============================] - 1s 27ms/step - loss: 0.1476 - acc: 0.9500 - val_loss: 0.1710 - val_acc: 0.9300 Epoch 9/30 40/40 [==============================] - 1s 27ms/step - loss: 0.1313 - acc: 0.9505 - val_loss: 0.1984 - val_acc: 0.9090 Epoch 10/30 40/40 [==============================] - 1s 27ms/step - loss: 0.1261 - acc: 0.9545 - val_loss: 0.1828 - val_acc: 0.9140 Epoch 11/30 40/40 [==============================] - 1s 31ms/step - loss: 0.1156 - acc: 0.9615 - val_loss: 0.1600 - val_acc: 0.9320 Epoch 12/30 40/40 [==============================] - 1s 27ms/step - loss: 0.1102 - acc: 0.9650 - val_loss: 0.1817 - val_acc: 0.9140 Epoch 13/30 40/40 [==============================] - 1s 27ms/step - loss: 0.1019 - acc: 0.9670 - val_loss: 0.1560 - val_acc: 0.9310 Epoch 14/30 40/40 [==============================] - 2s 38ms/step - loss: 0.0996 - acc: 0.9645 - val_loss: 0.1858 - val_acc: 0.9130 Epoch 15/30 40/40 [==============================] - 2s 40ms/step - loss: 0.0957 - acc: 0.9665 - val_loss: 0.1328 - val_acc: 0.9490 Epoch 16/30 40/40 [==============================] - 2s 42ms/step - loss: 0.0908 - acc: 0.9725 - val_loss: 0.1451 - val_acc: 0.9400 Epoch 17/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0852 - acc: 0.9715 - val_loss: 0.1453 - val_acc: 0.9370 Epoch 18/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0841 - acc: 0.9745 - val_loss: 0.1604 - val_acc: 0.9260 Epoch 19/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0762 - acc: 0.9760 - val_loss: 0.1258 - val_acc: 0.9540 Epoch 20/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0743 - acc: 0.9755 - val_loss: 0.1362 - val_acc: 0.9450 Epoch 21/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0703 - acc: 0.9785 - val_loss: 0.1283 - val_acc: 0.9450 Epoch 22/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0717 - acc: 0.9780 - val_loss: 0.1349 - val_acc: 0.9440 Epoch 23/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0656 - acc: 0.9820 - val_loss: 0.1375 - val_acc: 0.9410 Epoch 24/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0642 - acc: 0.9820 - val_loss: 0.1219 - val_acc: 0.9510 Epoch 25/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0606 - acc: 0.9850 - val_loss: 0.1105 - val_acc: 0.9520 Epoch 26/30 40/40 [==============================] - 2s 39ms/step - loss: 0.0611 - acc: 0.9820 - val_loss: 0.1189 - val_acc: 0.9510 Epoch 27/30 40/40 [==============================] - 2s 39ms/step - loss: 0.0579 - acc: 0.9820 - val_loss: 0.1279 - val_acc: 0.9450 Epoch 28/30 40/40 [==============================] - 2s 42ms/step - loss: 0.0563 - acc: 0.9860 - val_loss: 0.1362 - val_acc: 0.9420 Epoch 29/30 40/40 [==============================] - 1s 26ms/step - loss: 0.0516 - acc: 0.9855 - val_loss: 0.1468 - val_acc: 0.9370 Epoch 30/30 40/40 [==============================] - 1s 26ms/step - loss: 0.0507 - acc: 0.9865 - val_loss: 0.1306 - val_acc: 0.9450
import matplotlib
%matplotlib inline
import matplotlib.pyplot as plt
ent_acc = history2.history['acc']
val_acc = history2.history['val_acc']
ent_loss = history2.history['loss']
val_loss = history2.history['val_loss']
epochs = range(len(ent_acc))
plt.plot(epochs, ent_acc, 'ro', label='Entrenamiento')
plt.plot(epochs, val_acc, 'r', label='Validación')
plt.title('Accuracy en Entrenamiento y Validación')
plt.legend()
plt.figure()
plt.plot(epochs, ent_loss, 'ro', label='Entrenamiento')
plt.plot(epochs, val_loss, 'r', label='Validación')
plt.title('Pérdida en Entrenamiento y Validación')
plt.legend()
plt.show()
Para este tercer omodelo estamos usando EfficientNetB0, como ya he comentado antes es un modelo especializado en tratar con pocos elementos, en este caso he añadido un par de capas más que el ejemplo anterior.
Podemos ver en la gráfica que el modelo esta "anclado" no mejora con el paso de las épocas y el rendimiento no es bueno (65% y 50% respectivamente en entrenamiento y validación)
from keras import models
from keras import layers
from keras import optimizers
model3 = models.Sequential()
model3.add(layers.Dense(256, activation='relu'))
model3.add(layers.Dropout(0.5))
model3.add(layers.Dense(128, activation='relu'))
model3.add(layers.Dropout(0.5))
model3.add(layers.Dense(64, activation='relu'))
model3.add(layers.Dropout(0.5))
model3.add(layers.Dense(1, activation='sigmoid'))
model3.compile(optimizer=optimizers.RMSprop(lr=2e-5),
loss='binary_crossentropy',
metrics=['acc'])
history3 = model3.fit(train_features_Eff, train_labels_Eff,
epochs=20,
batch_size=40,
validation_data=(validation_features_Eff, validation_labels_Eff))
WARNING:absl:`lr` is deprecated in Keras optimizer, please use `learning_rate` or use the legacy optimizer, e.g.,tf.keras.optimizers.legacy.RMSprop.
Epoch 1/20 50/50 [==============================] - 8s 150ms/step - loss: 1.7258 - acc: 0.6060 - val_loss: 0.6953 - val_acc: 0.5080 Epoch 2/20 50/50 [==============================] - 6s 121ms/step - loss: 0.6672 - acc: 0.6340 - val_loss: 0.7045 - val_acc: 0.5080 Epoch 3/20 50/50 [==============================] - 9s 174ms/step - loss: 0.6598 - acc: 0.6340 - val_loss: 0.7167 - val_acc: 0.5080 Epoch 4/20 50/50 [==============================] - 7s 139ms/step - loss: 0.6575 - acc: 0.6340 - val_loss: 0.7219 - val_acc: 0.5080 Epoch 5/20 50/50 [==============================] - 6s 129ms/step - loss: 0.6576 - acc: 0.6340 - val_loss: 0.7266 - val_acc: 0.5080 Epoch 6/20 50/50 [==============================] - 5s 105ms/step - loss: 0.6679 - acc: 0.6320 - val_loss: 0.7248 - val_acc: 0.5080 Epoch 7/20 50/50 [==============================] - 9s 188ms/step - loss: 0.6571 - acc: 0.6340 - val_loss: 0.7234 - val_acc: 0.5080 Epoch 8/20 50/50 [==============================] - 6s 114ms/step - loss: 0.6588 - acc: 0.6340 - val_loss: 0.7217 - val_acc: 0.5080 Epoch 9/20 50/50 [==============================] - 5s 100ms/step - loss: 0.6587 - acc: 0.6340 - val_loss: 0.7224 - val_acc: 0.5080 Epoch 10/20 50/50 [==============================] - 6s 129ms/step - loss: 0.6574 - acc: 0.6340 - val_loss: 0.7248 - val_acc: 0.5080 Epoch 11/20 50/50 [==============================] - 7s 133ms/step - loss: 0.6583 - acc: 0.6340 - val_loss: 0.7233 - val_acc: 0.5080 Epoch 12/20 50/50 [==============================] - 6s 130ms/step - loss: 0.6596 - acc: 0.6340 - val_loss: 0.7203 - val_acc: 0.5080 Epoch 13/20 50/50 [==============================] - 5s 98ms/step - loss: 0.6585 - acc: 0.6340 - val_loss: 0.7234 - val_acc: 0.5080 Epoch 14/20 50/50 [==============================] - 6s 120ms/step - loss: 0.6566 - acc: 0.6340 - val_loss: 0.7256 - val_acc: 0.5080 Epoch 15/20 50/50 [==============================] - 8s 154ms/step - loss: 0.6581 - acc: 0.6340 - val_loss: 0.7225 - val_acc: 0.5080 Epoch 16/20 50/50 [==============================] - 6s 112ms/step - loss: 0.6568 - acc: 0.6340 - val_loss: 0.7261 - val_acc: 0.5080 Epoch 17/20 50/50 [==============================] - 8s 161ms/step - loss: 0.6592 - acc: 0.6340 - val_loss: 0.7234 - val_acc: 0.5080 Epoch 18/20 50/50 [==============================] - 7s 143ms/step - loss: 0.6571 - acc: 0.6340 - val_loss: 0.7270 - val_acc: 0.5080 Epoch 19/20 50/50 [==============================] - 8s 155ms/step - loss: 0.6581 - acc: 0.6340 - val_loss: 0.7244 - val_acc: 0.5080 Epoch 20/20 50/50 [==============================] - 5s 97ms/step - loss: 0.6587 - acc: 0.6340 - val_loss: 0.7240 - val_acc: 0.5080
import matplotlib
%matplotlib inline
import matplotlib.pyplot as plt
ent_acc = history3.history['acc']
val_acc = history3.history['val_acc']
ent_loss = history3.history['loss']
val_loss = history3.history['val_loss']
epochs = range(len(ent_acc))
plt.plot(epochs, ent_acc, 'bo', label='Entrenamiento')
plt.plot(epochs, val_acc, 'b', label='Validación')
plt.title('Accuracy en Entrenamiento y Validación')
plt.legend()
plt.figure()
plt.plot(epochs, ent_loss, 'bo', label='Entrenamiento')
plt.plot(epochs, val_loss, 'b', label='Validación')
plt.title('Pérdida en Entrenamiento y Validación')
plt.legend()
plt.show()
El mejor modelo que hemos encontrado ha sido el segundo (el que hemos usado VGG19) obteniendo alrededor de un 97% de acierto.
Vamos a aplicarle fine-tunning con muchas y pocas capas.
Aplicar fine-tunning consiste en descongelar algunas capas de la base preentrenada (VGG19) en nuestro caso y entrenar conjuntamente las descongeladas con las añadidas
conv_base_VGG19.summary()
Model: "vgg19"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_4 (InputLayer) [(None, 150, 150, 3)] 0
block1_conv1 (Conv2D) (None, 150, 150, 64) 1792
block1_conv2 (Conv2D) (None, 150, 150, 64) 36928
block1_pool (MaxPooling2D) (None, 75, 75, 64) 0
block2_conv1 (Conv2D) (None, 75, 75, 128) 73856
block2_conv2 (Conv2D) (None, 75, 75, 128) 147584
block2_pool (MaxPooling2D) (None, 37, 37, 128) 0
block3_conv1 (Conv2D) (None, 37, 37, 256) 295168
block3_conv2 (Conv2D) (None, 37, 37, 256) 590080
block3_conv3 (Conv2D) (None, 37, 37, 256) 590080
block3_conv4 (Conv2D) (None, 37, 37, 256) 590080
block3_pool (MaxPooling2D) (None, 18, 18, 256) 0
block4_conv1 (Conv2D) (None, 18, 18, 512) 1180160
block4_conv2 (Conv2D) (None, 18, 18, 512) 2359808
block4_conv3 (Conv2D) (None, 18, 18, 512) 2359808
block4_conv4 (Conv2D) (None, 18, 18, 512) 2359808
block4_pool (MaxPooling2D) (None, 9, 9, 512) 0
block5_conv1 (Conv2D) (None, 9, 9, 512) 2359808
block5_conv2 (Conv2D) (None, 9, 9, 512) 2359808
block5_conv3 (Conv2D) (None, 9, 9, 512) 2359808
block5_conv4 (Conv2D) (None, 9, 9, 512) 2359808
block5_pool (MaxPooling2D) (None, 4, 4, 512) 0
=================================================================
Total params: 20024384 (76.39 MB)
Trainable params: 20024384 (76.39 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
vamos a descongelar y entrenar desde la base block5 y ver como afecta al rendimiento
conv_base_VGG19.trainable = True
train = False
# recorre las capas y a partir de la block5_conv1 descongela
for layer in conv_base_VGG19.layers:
if layer.name == 'block5_conv1':
train = True
if train:
layer.trainable = True
else:
layer.trainable = False
model2.compile(optimizer=optimizers.RMSprop(learning_rate=2e-5),
loss='binary_crossentropy',
metrics=['acc'])
history2 = model2.fit(train_features_VGG19, train_labels_VGG19,
epochs=30,
batch_size=50,
validation_data=(validation_features_VGG19, validation_labels_VGG19))
Epoch 1/30 40/40 [==============================] - 2s 32ms/step - loss: 0.0496 - acc: 0.9870 - val_loss: 0.1195 - val_acc: 0.9500 Epoch 2/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0475 - acc: 0.9875 - val_loss: 0.1067 - val_acc: 0.9520 Epoch 3/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0433 - acc: 0.9900 - val_loss: 0.1007 - val_acc: 0.9560 Epoch 4/30 40/40 [==============================] - 2s 38ms/step - loss: 0.0432 - acc: 0.9900 - val_loss: 0.1144 - val_acc: 0.9520 Epoch 5/30 40/40 [==============================] - 2s 39ms/step - loss: 0.0426 - acc: 0.9895 - val_loss: 0.1099 - val_acc: 0.9510 Epoch 6/30 40/40 [==============================] - 2s 42ms/step - loss: 0.0415 - acc: 0.9890 - val_loss: 0.1016 - val_acc: 0.9530 Epoch 7/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0388 - acc: 0.9915 - val_loss: 0.1171 - val_acc: 0.9480 Epoch 8/30 40/40 [==============================] - 1s 31ms/step - loss: 0.0394 - acc: 0.9900 - val_loss: 0.1154 - val_acc: 0.9500 Epoch 9/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0352 - acc: 0.9940 - val_loss: 0.1009 - val_acc: 0.9540 Epoch 10/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0373 - acc: 0.9925 - val_loss: 0.1112 - val_acc: 0.9530 Epoch 11/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0360 - acc: 0.9915 - val_loss: 0.1041 - val_acc: 0.9520 Epoch 12/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0333 - acc: 0.9930 - val_loss: 0.1111 - val_acc: 0.9530 Epoch 13/30 40/40 [==============================] - 1s 28ms/step - loss: 0.0345 - acc: 0.9930 - val_loss: 0.0976 - val_acc: 0.9550 Epoch 14/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0311 - acc: 0.9940 - val_loss: 0.1025 - val_acc: 0.9540 Epoch 15/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0303 - acc: 0.9930 - val_loss: 0.1001 - val_acc: 0.9550 Epoch 16/30 40/40 [==============================] - 2s 39ms/step - loss: 0.0308 - acc: 0.9940 - val_loss: 0.1006 - val_acc: 0.9540 Epoch 17/30 40/40 [==============================] - 2s 41ms/step - loss: 0.0271 - acc: 0.9970 - val_loss: 0.1124 - val_acc: 0.9490 Epoch 18/30 40/40 [==============================] - 2s 42ms/step - loss: 0.0258 - acc: 0.9965 - val_loss: 0.0985 - val_acc: 0.9550 Epoch 19/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0238 - acc: 0.9965 - val_loss: 0.1006 - val_acc: 0.9560 Epoch 20/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0262 - acc: 0.9950 - val_loss: 0.0949 - val_acc: 0.9540 Epoch 21/30 40/40 [==============================] - 2s 41ms/step - loss: 0.0263 - acc: 0.9940 - val_loss: 0.1001 - val_acc: 0.9550 Epoch 22/30 40/40 [==============================] - 2s 45ms/step - loss: 0.0226 - acc: 0.9970 - val_loss: 0.0913 - val_acc: 0.9590 Epoch 23/30 40/40 [==============================] - 2s 42ms/step - loss: 0.0234 - acc: 0.9960 - val_loss: 0.0957 - val_acc: 0.9550 Epoch 24/30 40/40 [==============================] - 2s 45ms/step - loss: 0.0236 - acc: 0.9955 - val_loss: 0.1022 - val_acc: 0.9560 Epoch 25/30 40/40 [==============================] - 1s 35ms/step - loss: 0.0212 - acc: 0.9965 - val_loss: 0.0981 - val_acc: 0.9560 Epoch 26/30 40/40 [==============================] - 2s 60ms/step - loss: 0.0220 - acc: 0.9975 - val_loss: 0.0988 - val_acc: 0.9560 Epoch 27/30 40/40 [==============================] - 3s 74ms/step - loss: 0.0211 - acc: 0.9985 - val_loss: 0.0922 - val_acc: 0.9550 Epoch 28/30 40/40 [==============================] - 2s 50ms/step - loss: 0.0201 - acc: 0.9970 - val_loss: 0.1039 - val_acc: 0.9540 Epoch 29/30 40/40 [==============================] - 2s 55ms/step - loss: 0.0187 - acc: 0.9980 - val_loss: 0.0950 - val_acc: 0.9560 Epoch 30/30 40/40 [==============================] - 1s 29ms/step - loss: 0.0182 - acc: 0.9970 - val_loss: 0.0984 - val_acc: 0.9560
import matplotlib
%matplotlib inline
import matplotlib.pyplot as plt
ent_acc = history2.history['acc']
val_acc = history2.history['val_acc']
ent_loss = history2.history['loss']
val_loss = history2.history['val_loss']
epochs = range(len(ent_acc))
plt.plot(epochs, ent_acc, 'ro', label='Entrenamiento')
plt.plot(epochs, val_acc, 'r', label='Validación')
plt.title('Accuracy en Entrenamiento y Validación')
plt.legend()
plt.figure()
plt.plot(epochs, ent_loss, 'ro', label='Entrenamiento')
plt.plot(epochs, val_loss, 'r', label='Validación')
plt.title('Pérdida en Entrenamiento y Validación')
plt.legend()
<matplotlib.legend.Legend at 0x7f3ced0f7340>
En este primer caso podemos ver (aunque la gráfica no sea muy intuitiva) que obtenemos una leve mejora en el rendimiento y menos pérdida: Entrenamiento 98->99.8% Validación 95->95.6
Aquí hemos descongelado solo algunas de las últimas capas y las hemos entrenado con el resto del modelo que hemos creado,hemos obtenido una leve mejora puede ser debido a quede esta manera estamos haciendo que se adapte mejor a nuestro problema especificamente
conv_base_VGG19.trainable = True
train = False
for layer in conv_base_VGG19.layers:
if layer.name == 'block4_conv1':
train = True
if train:
layer.trainable = True
else:
layer.trainable = False
model2.compile(optimizer=optimizers.RMSprop(learning_rate=2e-5), # Corrige lr a learning_rate
loss='binary_crossentropy',
metrics=['acc'])
history2 = model2.fit(train_features_VGG19, train_labels_VGG19,
epochs=30,
batch_size=50,
validation_data=(validation_features_VGG19, validation_labels_VGG19))
Epoch 1/30 40/40 [==============================] - 2s 32ms/step - loss: 0.0201 - acc: 0.9960 - val_loss: 0.0929 - val_acc: 0.9570 Epoch 2/30 40/40 [==============================] - 1s 28ms/step - loss: 0.0161 - acc: 0.9980 - val_loss: 0.0982 - val_acc: 0.9570 Epoch 3/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0172 - acc: 0.9970 - val_loss: 0.1003 - val_acc: 0.9550 Epoch 4/30 40/40 [==============================] - 1s 27ms/step - loss: 0.0164 - acc: 0.9985 - val_loss: 0.1083 - val_acc: 0.9480 Epoch 5/30 40/40 [==============================] - 1s 31ms/step - loss: 0.0161 - acc: 0.9985 - val_loss: 0.0896 - val_acc: 0.9560 Epoch 6/30 40/40 [==============================] - 2s 58ms/step - loss: 0.0166 - acc: 0.9975 - val_loss: 0.1062 - val_acc: 0.9500 Epoch 7/30 40/40 [==============================] - 3s 63ms/step - loss: 0.0162 - acc: 0.9990 - val_loss: 0.1098 - val_acc: 0.9490 Epoch 8/30 40/40 [==============================] - 3s 68ms/step - loss: 0.0157 - acc: 0.9980 - val_loss: 0.1001 - val_acc: 0.9560 Epoch 9/30 40/40 [==============================] - 2s 52ms/step - loss: 0.0153 - acc: 0.9990 - val_loss: 0.0951 - val_acc: 0.9570 Epoch 10/30 40/40 [==============================] - 1s 35ms/step - loss: 0.0139 - acc: 0.9990 - val_loss: 0.0961 - val_acc: 0.9570 Epoch 11/30 40/40 [==============================] - 2s 41ms/step - loss: 0.0132 - acc: 0.9980 - val_loss: 0.1012 - val_acc: 0.9560 Epoch 12/30 40/40 [==============================] - 2s 45ms/step - loss: 0.0130 - acc: 0.9990 - val_loss: 0.0994 - val_acc: 0.9560 Epoch 13/30 40/40 [==============================] - 2s 52ms/step - loss: 0.0143 - acc: 0.9990 - val_loss: 0.0998 - val_acc: 0.9560 Epoch 14/30 40/40 [==============================] - 2s 59ms/step - loss: 0.0120 - acc: 0.9990 - val_loss: 0.1051 - val_acc: 0.9560 Epoch 15/30 40/40 [==============================] - 2s 51ms/step - loss: 0.0121 - acc: 1.0000 - val_loss: 0.0946 - val_acc: 0.9560 Epoch 16/30 40/40 [==============================] - 2s 43ms/step - loss: 0.0113 - acc: 0.9990 - val_loss: 0.0975 - val_acc: 0.9570 Epoch 17/30 40/40 [==============================] - 2s 48ms/step - loss: 0.0112 - acc: 0.9995 - val_loss: 0.1019 - val_acc: 0.9560 Epoch 18/30 40/40 [==============================] - 1s 35ms/step - loss: 0.0109 - acc: 1.0000 - val_loss: 0.0973 - val_acc: 0.9580 Epoch 19/30 40/40 [==============================] - 1s 35ms/step - loss: 0.0104 - acc: 0.9990 - val_loss: 0.1111 - val_acc: 0.9510 Epoch 20/30 40/40 [==============================] - 2s 40ms/step - loss: 0.0101 - acc: 0.9995 - val_loss: 0.1035 - val_acc: 0.9560 Epoch 21/30 40/40 [==============================] - 2s 54ms/step - loss: 0.0102 - acc: 0.9990 - val_loss: 0.0903 - val_acc: 0.9580 Epoch 22/30 40/40 [==============================] - 2s 61ms/step - loss: 0.0100 - acc: 1.0000 - val_loss: 0.1104 - val_acc: 0.9500 Epoch 23/30 40/40 [==============================] - 2s 55ms/step - loss: 0.0107 - acc: 0.9990 - val_loss: 0.0964 - val_acc: 0.9570 Epoch 24/30 40/40 [==============================] - 1s 37ms/step - loss: 0.0093 - acc: 1.0000 - val_loss: 0.0955 - val_acc: 0.9580 Epoch 25/30 40/40 [==============================] - 1s 37ms/step - loss: 0.0099 - acc: 0.9995 - val_loss: 0.0948 - val_acc: 0.9570 Epoch 26/30 40/40 [==============================] - 1s 32ms/step - loss: 0.0081 - acc: 1.0000 - val_loss: 0.0980 - val_acc: 0.9570 Epoch 27/30 40/40 [==============================] - 1s 35ms/step - loss: 0.0092 - acc: 1.0000 - val_loss: 0.0944 - val_acc: 0.9580 Epoch 28/30 40/40 [==============================] - 1s 34ms/step - loss: 0.0083 - acc: 1.0000 - val_loss: 0.0969 - val_acc: 0.9580 Epoch 29/30 40/40 [==============================] - 1s 33ms/step - loss: 0.0076 - acc: 1.0000 - val_loss: 0.0968 - val_acc: 0.9580 Epoch 30/30 40/40 [==============================] - 2s 59ms/step - loss: 0.0076 - acc: 1.0000 - val_loss: 0.1032 - val_acc: 0.9560
import matplotlib
%matplotlib inline
import matplotlib.pyplot as plt
ent_acc = history2.history['acc']
val_acc = history2.history['val_acc']
ent_loss = history2.history['loss']
val_loss = history2.history['val_loss']
epochs = range(len(ent_acc))
plt.plot(epochs, ent_acc, 'ro', label='Entrenamiento')
plt.plot(epochs, val_acc, 'r', label='Validación')
plt.title('Accuracy en Entrenamiento y Validación')
plt.legend()
plt.figure()
plt.plot(epochs, ent_loss, 'ro', label='Entrenamiento')
plt.plot(epochs, val_loss, 'r', label='Validación')
plt.title('Pérdida en Entrenamiento y Validación')
plt.legend()
<matplotlib.legend.Legend at 0x7f3cecd0d180>
Podemos suavizar un poco la curva como con el metodo smooth_curve visto en clase en la práctica 5
def smooth_curve(points, factor=0.8):
smoothed_points = []
for point in points:
if smoothed_points:
previous = smoothed_points[-1]
smoothed_points.append(previous * factor + point * (1 - factor))
else:
smoothed_points.append(point)
return smoothed_points
plt.plot(epochs,
smooth_curve(ent_acc), 'bo', label='Acc Entrenamiento Suavizada')
plt.plot(epochs,
smooth_curve(val_acc), 'b', label='Acc Validación Suavizada')
plt.title('Accuracy en Entrenamiento y Validación')
plt.legend()
plt.figure()
plt.plot(epochs,
smooth_curve(ent_loss), 'bo', label='Pérdida Entrenamiento Suavizada')
plt.plot(epochs,
smooth_curve(val_loss), 'b', label='Pérdida Validación Suavizada')
plt.title('Pérdida en Entrenamiento y Validación')
plt.legend()
plt.show()
En este segundo caso noto que el modelo se sobreajusta un poco más, ya que en el entrenamiento alcanza el 100% mientrás que en validación empeora minimamente. Quizás al descongelar y entrenar tantas capas he provocado que el modelo se adapte demasiado a mi modelo y provoque ese sobreajuste, o quizás al no ser el modelo muy grande con tantas capas provoca también el sobreajuste.
Vamos a comprobar varios valores en concreto aaccuracy,precision,recall y f1 asi que voy a comentar brevemente que nos da esa información.
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
predictions = model.predict(test_features)
# Convertimos en etiquetas binarias
rounded_predictions = np.round(predictions)
# F1-score
f1 = f1_score(test_labels, rounded_predictions)
recall = recall_score(test_labels, rounded_predictions)
# Evaluar el modelo
test_loss, test_acc = model.evaluate(test_features, test_labels)
print('Test accuracy:', test_acc)
print('test_loss:', test_loss)
print('Test F1-score:', f1)
print('Recall:', recall)
32/32 [==============================] - 1s 32ms/step 32/32 [==============================] - 1s 36ms/step - loss: 0.2742 - acc: 0.9640 Test accuracy: 0.9639999866485596 test_loss: 0.2742430567741394 Test F1-score: 0.96709323583181 Recall: 0.9742173112338858
Podemos ver que para el modelo 1 hemos obtenido una precisión del 96.4% teniendo en cuenta que en validación obtuvimos un 96.7 % podemos decir que es un buen rendimiento para el test, ya que es próximo al rendimiento obtenido en los otros casos. Como añadido podemos ver que la perdida es baja (indicativo de un buen modelo). Para los valores de f1 y recall al ser próximos a 1 nos indica que el modelo es bueno clasificando
predictions2 = model2.predict(test_features_VGG19)
rounded_predictions = np.round(predictions2)
# F1-score
f1 = f1_score(test_labels, rounded_predictions)
recall = recall_score(test_labels, rounded_predictions)
# Evaluar el modelo
test_loss, test_acc = model2.evaluate(test_features_VGG19, test_labels_VGG19)
print('Test accuracy:', test_acc)
print('test_loss:', test_loss)
print('Test F1-score:', f1)
print('Recall:', recall)
32/32 [==============================] - 1s 15ms/step 32/32 [==============================] - 1s 11ms/step - loss: 0.0900 - acc: 0.9630 Test accuracy: 0.9629999995231628 test_loss: 0.09000647068023682 Test F1-score: 0.5459558823529412 Recall: 0.5469613259668509
Para nuestro segundo modelo obtenemos un rendimiento similar, la precisión alrededor de un 96%, aunque en este caso la pérdida es menor,la pérdida es una medida de cuánto se desvían las predicciones del modelo de las etiquetas reales en el conjunto de pruebas. Aunque en este caso las métricas del F1 y el recall son más bajas
predictions3 = model3.predict(test_features_Eff)
rounded_predictions = np.round(predictions3)
# Calcular el F1-score
f1 = f1_score(test_labels, rounded_predictions)
recall = recall_score(test_labels, rounded_predictions)
# Evaluar el modelo
test_loss, test_acc = model3.evaluate(test_features_Eff, test_labels_Eff)
print('Test accuracy:', test_acc)
print('test_loss:', test_loss)
print('Test F1-score:', f1)
print('Recall:', recall)
32/32 [==============================] - 1s 30ms/step 32/32 [==============================] - 1s 18ms/step - loss: 0.7197 - acc: 0.5160 Test accuracy: 0.515999972820282 test_loss: 0.7197336554527283 Test F1-score: 0.7038237200259235 Recall: 1.0
Ahora para el conjunto nuevo haremos lo mismo que al principio, de drive obtendremos las imagenes de un subconjunto que he obtenido(el del enlace de la entrega que recomiendan usar) y he reducido su tamaño manteniendo la proporción ya que era muy grande
from google.colab import drive
drive.mount('/content/drive')
Mounted at /content/drive
# Copiamos el dataset y lo decomprimimos
!cp /content/drive/MyDrive/Face_Mask_Dataset.zip .
!unzip -q Face_Mask_Dataset.zip
# Desmontamos
drive.flush_and_unmount()
import os
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
base_dir = 'Face_Mask_Dataset'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
test_dir = os.path.join(base_dir, 'test')
datagen = ImageDataGenerator(rescale=1./255)
batch_size = 20
def extract_features2(directory, sample_count):
features = np.zeros(shape=(sample_count, 4, 4, 512))
labels = np.zeros(shape=(sample_count))
generator = datagen.flow_from_directory(
directory,
target_size=(150, 150),
batch_size=batch_size,
class_mode='binary')
i = 0
for inputs_batch, labels_batch in generator:
features_batch = conv_base_VGG16.predict(inputs_batch)
features[i * batch_size : (i + 1) * batch_size] = features_batch
labels[i * batch_size : (i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
break
return features, labels
test_features, test_labels = extract_features2(test_dir, 1000)
Found 2073 images belonging to 2 classes. 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 8s 8s/step 1/1 [==============================] - 12s 12s/step 1/1 [==============================] - 8s 8s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 6s 6s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 5s 5s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step 1/1 [==============================] - 4s 4s/step
test_features=np.reshape(test_features, (1000, 4 * 4 * 512))
Ahora que hemos extraido las características vamos a comprobar las métricas anteriores para este nuevo conjunto
Prueba con el modelo 1: VGG16
predictions = model.predict(test_features)
# Convertimos las probabilidades en etiquetas binarias
rounded_predictions = np.round(predictions)
f1 = f1_score(test_labels, rounded_predictions)
recall = recall_score(test_labels, rounded_predictions)
# Evaluar el modelo original
test_loss, test_acc = model.evaluate(test_features, test_labels)
print('Test accuracy:', test_acc)
print('Test loss:', test_loss)
print('Test F1-score:', f1)
print('Recall:', recall)
32/32 [==============================] - 0s 10ms/step 32/32 [==============================] - 0s 11ms/step - loss: 1.0188 - acc: 0.9150 Test accuracy: 0.9150000214576721 Test loss: 1.0188205242156982 Test F1-score: 0.9177153920619554 Recall: 0.9978947368421053
Para este caso vemos que aunque la pérdida sea algo mayor obtenemos un buen rendimiento (91.5%). Además los valores de f1 y recall son bastante mejores que en el caso anterior, en general, se nos muestra que para este nuevo conjunto de mascarillas el modelo generado da un buen resultado
Prueba con VGG19
predictions = model.predict(test_features)
# Convertir las probabilidades en etiquetas binarias
rounded_predictions = np.round(predictions)
f1 = f1_score(test_labels, rounded_predictions)
recall = recall_score(test_labels, rounded_predictions)
# Evaluar el modelo original
test_loss, test_acc = model2.evaluate(test_features, test_labels)
print('Test accuracy:', test_acc)
print('Test loss:', test_loss)
print('Test F1-score:', f1)
print('Recall:', recall)
32/32 [==============================] - 1s 25ms/step 32/32 [==============================] - 1s 14ms/step - loss: 0.3543 - acc: 0.8520 Test accuracy: 0.8519999980926514 Test loss: 0.35433080792427063 Test F1-score: 0.9177153920619554 Recall: 0.9978947368421053
Para este segundo modelo el rendimiento ha decaido hasta un 85%, aunque tiene un f1 y recall , indicando que cuando clasifica bien una imagen como positiva(mascarilla) esta bastante seguro de la fiabilidad de esta